Playbook 01 · Tested Oct 14, 2026 · Original workflow
The cited company brief
From a company name to a sourced, one-page investor brief in well under an hour, with every number traced back to a filing.
Who it's for
- VC: a fast, sourced read on a public comp or category leader before a partner meeting or a market map.
- PE: a first-pass brief on a target's listed peers before an IC pre-read, or a refresher on a portfolio company's competitor.
- Public markets: a dated, cited snapshot before an earnings call or when picking up a name for coverage.
The workflow is about speed with an audit trail. It replaces the first draft, not your judgment. It works best on companies with SEC filings. For private companies, swap in the company's own disclosures, court and regulatory filings and reputable press, and expect more gaps.
The time it takes
The tools
Claude (Claude Code CLI)Research with web search, then the draft · in our directoryCCodex CLI ↗Independent second-model fact checkSSEC EDGAR ↗Primary filings and a final recency checkWe ran Claude Code and Codex from the command line on existing subscriptions; no API keys and no paid data. The chat apps work too: paste the same prompts into Claude or ChatGPT with web search turned on for step 1 and off for step 2. Research platforms in our directory that cover parts of this workflow: Perplexity Finance, ChatGPT, AlphaSense, Hebbia, Rogo, Daloopa.
Step by step, with the exact prompts
Build a source pack, not a brief3 min 5 s
Claude Code with web search and fetch
Ask for primary sources first (10-K, earnings release, monthly sales), then at most three reputable secondary sources. Ask for verbatim quotes per source and a "Gaps" list. Keeping research separate from writing means the writer can only use what was actually found.
Show the exact prompt
You are a research analyst preparing a source pack on Costco Wholesale Corporation (NASDAQ: COST) for an investor. Today is October 14, 2026. Task: build a SOURCE PACK, not a brief yet. 1. Use web search and fetch to find primary sources first: Costco's most recent Form 10-K (fiscal year ended around August 30, 2026, or FY2025 if FY2026 is not yet filed), the most recent quarterly/annual earnings press release, Costco's monthly sales release(s), and investor.costco.com. Then add at most three reputable secondary sources (e.g. Reuters, AP, WSJ, CNBC) for recent events. 2. For each source record: ID (S1, S2...), title, publisher, date, exact URL, and 3-8 VERBATIM quotes (copy text exactly as it appears) that contain the facts an investor needs: revenue, net income, membership fee income, member counts / renewal rates, warehouse count, comparable sales, business model, risks, capital return (dividends/buybacks), recent strategic events, leadership. 3. Do not paraphrase inside quotes. If you could not open a page, say so and do not quote it. Never invent a URL. Output Markdown only, with one section per source. End with a short list "Gaps" of facts you looked for but could not find in a primary source.
What happened in our run: Claude found the FY26 10-K, the Q4 release, the September sales release and one FOX Business story. It flagged its own cut-off quotes, and it reported CNBC and investor.costco.com as unreadable instead of guessing.
Draft from the pack only54 s
Claude Code, tools turned off
A fresh session with no web access writes the brief from the pack alone, with an [S#] citation on every factual sentence. Turning tools off stops the model from topping up with half-remembered figures.
Show the exact prompt
Using ONLY the source pack below, write a one-page investor brief on Costco Wholesale (COST). Rules: - Maximum ~450 words of body text. Sections: Snapshot (one paragraph), Business model, Latest results (a small table is fine), Members and unit economics, Capital return, What to watch (3 bullets: risks and open questions), Sources. - Every factual sentence or table cell ends with an inline citation like [S2]. Numbers must match the quote exactly (you may round, and say so: e.g. "about $X billion"). - If a fact is not in the source pack, leave it out. Do not use your own memory for any figure, date or name. - No valuation opinion, no price targets, no buy/sell language. Neutral, plain English. - Under Sources list each S# with title, publisher, date and URL exactly as in the pack. SOURCE PACK: [paste the source pack from step 1]
What happened in our run: The draft was tidy and well cited. Its own "check before publishing" list was a useful first triage.
Cross-check with a second model1 min 18 s
Codex CLI, read-only
A different model fact-checks every claim against the pack and returns a claim-by-claim table with verdicts and suggested fixes. Different models make different mistakes, which is the point.
Show the exact prompt
You are a fact-checker. Below is an investor brief with inline citations [S#] followed by the source pack it was written from. For EVERY factual claim in the brief (each sentence, bullet and table cell), output a row in a Markdown table: | # | Claim (short) | Cited | Supporting quote (verbatim from the pack, or "NONE") | Verdict | Verdicts: SUPPORTED (quote states it), ROUNDED-OK (rounding is faithful), DERIVED (correct arithmetic from quotes; show the math), UNSUPPORTED (no quote supports it), WRONG-SOURCE (supported only by a different S#), CONTRADICTED. Be strict. Then list any claims needing a fix and the exact corrected wording. BRIEF: [paste the brief from step 2] SOURCE PACK: [paste the source pack from step 1]
What happened in our run: Codex rated 94 claims and found five real problems: a mislabeled metric, an overstated cause, a wrong-source attribution, an ambiguous Labor Day effect, and an "n/a" that should be "not reported".
Check the claims against the originals yourselfabout 10 min
EDGAR, a browser, and text search
Download the filings and search each claim's number or phrase in the original. Fix anything the model mislabeled, and check the company's filing list for anything filed in the last week. This is the step you can't skip; the two models only checked each other against the same pack.
What happened in our run: We checked all 23 claims against the original filings or article, restored a cut-off quote, and added three facts the pack missed, including a dividend declared that morning.
Final edit and checking notesabout 2 min
You
Apply the fixes. Keep the checking table with the brief so the next reader can audit it in minutes. Label secondary-source claims as such.
The finished brief
This is the real output after checking, unedited apart from the fixes listed below. Each chip links to its source.
Costco Wholesale Corporation (Nasdaq: COST)
As of October 14, 2026. Fiscal 2026 = 52 weeks ended August 30, 2026. $ in USD.
Snapshot
Costco runs membership warehouses and e-commerce sites S1. At August 30, 2026 it operated 939 warehouses: 647 in the U.S. and Puerto Rico, 115 in Canada, and the rest across 12 other countries S1S2. It ended the year with 355,000 employees S1. Fiscal 2026 net sales rose 10.1% to $297.2 billion, and net income was $9.226 billion, or $20.76 per diluted share, against $18.21 a year earlier S2. Ron M. Vachris is President and CEO; Gary Millerchip is EVP and CFO S1.
Business model
Costco's stated concept is that "offering low prices on a limited selection of nationally-branded and private-label products in a wide range of categories will produce high sales volumes and rapid inventory turnover" S1. Membership fees were $5.907 billion of $303.154 billion in total revenue, under 2%, but equal to about half of the $11.685 billion operating income S2. Executive members, who pay more for a reward on purchases, drove about 75.2% of worldwide net sales S1.
Latest results
| Q4 FY26 (16 wks) | FY26 (52 wks) | Sept 2026 (5 wks) | |
|---|---|---|---|
| Net sales | $93.9B, +11.2% S2 | $297.2B, +10.1% S2 | $30.02B, +13.0% S3 |
| Comparable sales, total | 9.4% S2 | 8.4% S2 | 11.4% S3 |
| Comparable sales, ex gas and FX | 6.7% S2 | 6.6% S2 | 7.6% S3 |
| Digitally-enabled comparable sales | 19.5% S2 | 20.9% S2 | — |
| Digitally-enabled sales growth | — | — | 19.0% S3 |
| Net income / diluted EPS | $2.998B / $6.75 S2 | $9.226B / $20.76 S2 | Not reported monthly |
Q4 EPS included a non-recurring benefit of $0.15 per diluted share from IEEPA tariff refunds, net of partial reinvestment "in increased member values" S2.
A later Labor Day lifted September total and comparable sales by "a little more than 50bps" S3.
Members and unit growth
Paid members were 84.1 million, up from 81.0 million in fiscal 2025 and 76.2 million in fiscal 2024; 42.3 million were Executive members; total cardholders were 150.4 million S1. The renewal rate was 92.3% in the U.S. and Canada and 89.8% worldwide, a trailing measure S1. Membership fee revenue rose 11%, driven by new sign-ups, the fee increase in the U.S. and Canada effective September 1, 2024, and upgrades to Executive membership. Costco says "the fee income increase accounted for approximately 30% and 40% of membership income growth during 2026 and 2025" S1. It opened 28 warehouses in fiscal 2026, including three relocations, for 25 net new, and plans up to 33 openings in fiscal 2027, including five relocations S1.
Capital return
Dividends declared in fiscal 2026 were $2.458 billion ($5.54 per share), against $2.183 billion ($4.92) in fiscal 2025. The quarterly rate rose from $1.30 to $1.47 in April 2026 S1. On October 14, 2026 the board declared $1.47 again, payable November 13 to holders of record on October 30 S5. Costco bought back 891,000 shares at an average $950.90, for $847 million; $1.115 billion was left under a $4 billion authorization that expires in January 2027 S1.
What to watch
- Tariffs. The 10-K warns that "higher tariffs could adversely impact our results" S1. On the earnings call, as reported by FOX Business, Costco said it received $184 million in Q4 tariff refunds ($174 million plus $10 million in interest), reinvested most of it in lower prices, and intends to keep reinvesting the majority of future refunds S4. Costco's own release quantifies only the $0.15 per share S2.
- Concentration. U.S. and Canadian operations made up 86% of net sales and 83% of operating income in fiscal 2026, and Costco names strong competition from other retailers and warehouse clubs as a risk S1.
- Underlying growth. September comparable sales ex gas and FX were 7.6%. The Labor Day lift applies to total and comparable sales; the release doesn't give its effect on the adjusted figure S3.
Sources
- S1Costco Wholesale Corp., Form 10-K for the fiscal year ended August 30, 2026 · SEC EDGAR · Filed Oct 7, 2026
- S2Costco Wholesale Corporation Reports Fourth Quarter and Fiscal Year 2026 Operating Results (8-K, Ex. 99.1) · SEC EDGAR · Sep 24, 2026
- S3Costco Wholesale Corporation Reports September Sales Results · Costco investor relations (GlobeNewswire) · Oct 7, 2026
- S4Costco got $184 million in tariff refunds — here's what it did with the money · FOX Business (Eric Revell), via FOX 29 · Sep 26, 2026
- S5Costco Wholesale Corporation Announces Quarterly Cash Dividend (8-K, Ex. 99.1) · SEC EDGAR · Oct 14, 2026
Descriptive only. No valuation view, no recommendation.
Checking notes
Every claim in the brief, checked against the original document, not against the AI's quotes. Of 23 claims: 17 verified (one by arithmetic), 4 corrected or rephrased, 1 added from a filing the AI missed, and 1 verified against a secondary source only. Two of the verified rows (Executive-member sales share and U.S./Canada concentration) were added from the 10-K during the human check.
| Claim | Where we checked | Result | Note |
|---|---|---|---|
| 939 warehouses at Aug 30, 2026 | S1, Item 2 Properties | Verified | |
| 647 U.S. and Puerto Rico, 115 Canada, 12 other countries | S2, body text | Verified | Country count done by hand: 939 − 647 − 115 = 177 across 12 listed countries, matching the 10-K's 177 Other International. |
| 355,000 employees | S1, Item 1 Employee Base | Verified | |
| FY26 net sales +10.1% to $297.2B; net income $9.226B; EPS $20.76 vs $18.21 | S2, body and income statement; S1 MD&A | Verified | The 10-K rounds growth to 10%; the release gives 10.1%. |
| CEO and CFO names and titles | S1, Executive Officers | Verified | |
| Business-model quote | S1, Item 1 General | Verified verbatim | |
| Membership fees under 2% of revenue but about half of operating income | S2, income statement | Derived | 5,907 ÷ 303,154 = 1.95% of revenue; 5,907 ÷ 11,685 = 50.6% of operating income. This is arithmetic, not a Costco statement. |
| Executive members about 75.2% of net sales | S1, Item 1 Membership | Verified | Added during the human check; not in the AI draft. |
| Q4, FY and September results table | S2 tables; S3 PDF | Verified | Every cell was checked against the original table. |
| Digitally-enabled comparable sales 19.5% and 20.9% | S2, comparable sales table | Corrected | The AI draft labeled these as "sales growth". They're comparable sales. Codex flagged it, and the table now splits the two metrics. |
| $0.15 tariff-refund benefit, net of reinvestment | S2, body text | Corrected | The draft said refunds went into "member prices". S2 says "member values", so we used S2's wording. |
| Labor Day lift of a little more than 50bps | S3 PDF | Verified | Scope clarified: total and comparable sales only. |
| Members 84.1M / 81.0M / 76.2M; Executive 42.3M; cardholders 150.4M | S1, Item 1 membership table | Verified | Unit check: the table is "in thousands", so 84,100 = 84.1 million. |
| Renewal rates 92.3% and 89.8% | S1, Item 1 and MD&A | Verified | Costco defines this as a trailing calculation (renewals 7 to 18 months before the reporting date). |
| Fee revenue +11% and its three drivers | S1, MD&A highlights | Corrected | The fetch tool cut the quote off, so the AI pack had two drivers. The original lists three; we restored upgrades to Executive membership. |
| 30% / 40% of membership income growth | S1, MD&A Membership Fees | Kept as a direct quote | The draft paraphrased this as the fee hike causing the growth. Codex rated that an overreach, so we quote Costco's wording instead. |
| 28 openings / 25 net; up to 33 in FY27 | S1, MD&A highlights and Liquidity | Verified | |
| Dividends $2.458B / $5.54 vs $2.183B / $4.92; $1.30 to $1.47 | S1, MD&A Dividends and Note 6 | Verified | |
| Oct 14, 2026 dividend declaration | S5 | Added | Filed the day we ran this. The AI research step missed it, and a manual EDGAR check found it. |
| Buybacks 891,000 shares at $950.90 = $847M; $1.115B left of $4B | S1, Item 5 and MD&A | Verified | Arithmetic check: 891,000 × $950.90 = $847.2M. |
| "Higher tariffs could adversely impact our results" | S1 | Verified verbatim | |
| $184M Q4 refunds ($174M + $10M interest), mostly reinvested; CFO intends to continue | S4 | Verified against secondary source only | This matches the FOX Business article. We didn't check it against a primary call transcript, so the brief attributes it to FOX. |
| 86% / 83% of net sales and operating income from U.S. and Canada; competition risk | S1, Item 1A | Verified | Added during the human check. |
Left out on purpose: analyst price targets and "membership missed expectations" (only seen in search snippets; CNBC returned 403), and the "one-third of expected refunds" remark (its referent is ambiguous in the article).
Common failure modes, and how to catch them
Invented or wrong citations
The model cites a source that doesn't exist, or a real source that doesn't say what's claimed.
Check: Open every link. Search the original for the exact number, not the model's paraphrase. Prefer EDGAR links with an accession number over aggregator pages.
In our run: no invented URLs. One path used a zero-padded CIK; we changed it to the standard form.
Cut-off or stitched quotes
Web fetch tools often cut text off or stitch fragments together, so a quote can lose a clause that changes its meaning.
Check: Treat any "…" in a quote as unverified until you've read the full sentence in the original.
In our run: the membership-fee drivers lost "upgrades to Executive membership". We restored it.
Mislabeled metrics
A number lands under the wrong label, such as comparable sales vs. total sales growth, or GAAP vs. adjusted.
Check: Check the row and column headers in the source table, not just the number.
In our run: digitally-enabled comparable sales were labeled as sales growth. Codex caught it.
Causation creep
"X accounted for 30% of growth" becomes "X drove growth" in the summary.
Check: When attribution matters, quote the company's sentence instead of paraphrasing it.
In our run: the draft overstated the fee-increase attribution. We now quote it.
Unit and period errors
Thousands read as millions, 16-week quarters compared with 13-week ones, fiscal and calendar years mixed up.
Check: Read the table heading ("in thousands", "16 weeks ended") and write the period into every column header.
In our run: correct, but only because the pack carried the "in thousands" note.
Recency gaps
Search indexes lag, so a filing from this morning won't show up.
Check: Finish with a direct look at the company's EDGAR filing list for the last 7 days.
In our run: the AI missed a dividend declaration filed the same day. We added it from EDGAR.
Secondary-source drift
News stories round, merge or reword what management said on a call.
Check: Attribute call remarks to the outlet in the brief, or check them against a transcript.
In our run: $184 million included $10 million of interest. Our brief says so and attributes it to FOX Business.
Search snippets passed off as sources
Paywalled or blocked pages still show snippets, such as analyst targets or "missed expectations".
Check: If you couldn't open it, you can't cite it. Leave it out.
In our run: CNBC returned 403, so analyst views were left out.
Make it yours
- Add your own house sections (for example, a "questions for management" list), but keep the rule that every sentence carries a citation.
- For a portfolio or watchlist, run step 1 per company and save the source packs. Next quarter you only refresh the pack.
- Don't put confidential material into tools your firm hasn't approved. This workflow uses public sources only.
This playbook shows a research workflow. The company is an example, chosen because its filings are public and well covered. Nothing here is investment advice or a view on the stock. Figures are as of the dates shown and will go stale.